Author | Nanfengchuang reporter Zhu Qiuyu
"Hello everyone, I am Gao Yixiang godfrey. Although I have left this world, my heart will always be with you. I have seen your messages, you are my most important Precious treasure, I love you forever."
html In March, a man named "Gao Yixiang" in a suit, wearing a watch and a slicked back hair appeared on social media with a Taiwanese accent. Joining him in expressing his longing for fans in front of people's eyes are the late stars Coco Lee, Qiao Renliang, Kobe Bryant, Leslie Cheung... The only difference withis that the producer specifically noted on the side: "The video and sound are both generated by AI. ”
A few people expressed their weird feelings under the video. Fans of Coco Lee pointed out that the voice and appearance of ai Coco Lee are very similar to the original version, but there are still big differences in accent, pauses, and speaking emotions. "This is not Coco or similar to Coco at all. She is just a virtual image with a shell and an unknown person at its core. It may even be developed into a commodity." The resurrection technology of
ai has indeed become an industry chain. On various short video platforms, many bloggers sell AI resurrection technology in the name of technology for good. Many bloggers quoted prices to reporters:
makes the deceased's photos move and speak, 298 yuan;
clones the deceased's voice and video, 1498 yuan...
has been tested by Nanfengchuang reporters, and the currently popular AI voice cloning and digital The threshold for people with technical skills is not high, and there are many very mature open source projects on the market. Servers can easily "resurrect" multiple people every day at a low cost.
The deeper contradictions are exposed by celebrity parents. On March 16, Qiao Renliang's father told the media that he had seen the image of "his son being resurrected" spread online and "cannot accept it and feels uncomfortable." "They didn't ask for our consent. It was my niece who found the video and sent it to me. This is exposing scars."
The entanglement between technology and humanities was put in front of us. Is this a technology that improves people's ability to deal with the trauma of death, or is it distorting our sense of reality?
01
Resurrection of Stars
ai Many people learned about the power of resurrection from a celebrity father, the musician Bao Xiaobai.
html In March, when he was accepting the media, he showed the "appearance" of his daughter being resurrected by AI. Whenfaced reporters from many media, he first said to his mobile phone: "Bao Xiaorong, I am being interviewed now. Say hello to your friends from the media." The "tolerance" in the
mobile app has been waiting for a while. Finally, he said: "Hello, friends from the media, nice to meet you."
According to Bao Xiaobai, this AI "tolerance" is a public welfare project between him and his friend Liu Yan. A virtual character trained from the memories of my daughter’s 20 years before her death. After eight months of hard work, "Tolerance" finally returned to Bao Xiaobai, and he could talk, sing and interact with people in real time.
Bao Xiaobai was deeply touched by the experience of his daughter's successful "resurrection". He established a company called "Love Language Inclusion" and stated externally: "For a long time, AI has been negative to the public, either defrauding or replacing human work. But AI can also have good uses."
Him He believes that AI resurrection technology can become a kind of companionship. "Even if everyone knows it is fake, they will still accept it happily."
's similar "soul comfort" was circulated on the Internet many months after the death of superstar Coco Lee. On March 13, a lively Coco Lee appeared in front of people wearing a blue denim top and braided side braids. She said the above-mentioned "Gao Yixiang" words to the camera, which aroused the creative desire of many AI bloggers.
html In March, a blogger who claimed that "my wife is a fan of Coco Lee" recorded the process of resurrecting Coco Lee on station B. He first screened Coco Lee's original voice on the Internet, processed it, and put the clean voice into the AI model. He believed that after a night of calculations and training, more than 95% of Coco Lee's voice had been restored. After thevideo was released, he received many reminders from fans. He only considered Coco Lee's voice and appearance, "but in fact, I didn't consider too much the tone and mantra she used."The blogger
then made his second creation. This time, when cloning Li Min, he focused on marking the mantras and "establishing a coco mantra model." After nearly 4 days and 1,000 rounds of training each time, he finally cloned Coco Lee’s voice model. From this, he created a 4-minute voice message given to fans by AI Coco Lee.
"I am Coco Li Min, long time no see, I know you miss me."
"In China, there is a saying, 'There is no such thing as a banquet that lasts forever'. Maybe each of us is a passer-by in each other's lives, but I prefer a sentence in English, 'everything happened is good'. Everything that happens has a good side, so for me, the best thing is to stop at the moment when you are in love, I really I feel very happy."
After hearing these words, many people shed tears and praised him for using AI to create dreams.
But the most liked comments below represent the dissatisfaction of some fans with this approach. "I support the creation of the works of deceased artists, but I am opposed to imitating their personalities. This is too offensive."
"ai coco keeps saying thank you to her fans. She is very happy, but wouldn't the real her be when she is confused? Pain or despair? Who can really copy another person's personality?"
02
Near zero cost
The controversy itself does not stem from technology. However, the technical means and implementation subjects that can be included in "AI resurrection" are very diverse, which makes people have complex and confusing emotions in the face of technological progress. The situation of
package Xiaobai is more special. When the celebrity father spent so much of his energy resurrecting his daughter, they were pursuing a digital persona that resembled inclusion. This requires relatives to reconstruct and record the life nodes, personality, and lifetime images of the deceased, which relies on a huge amount of data.
But if an outsider resurrects a star, it will be much easier. Many AI resurrection packages on the market are also sold in the same way - hand over the images, sounds or photos of your loved ones during their lifetime to AI, and let AI train a voice model or digital person. Then, the consumer can provide a paragraph of words that the AI relatives want to say and let the AI say it.
html This is how Zhang Lin, who was born in 090, found an AI to resurrect an internet celebrity blogger during the Spring Festival this year, hoping to resurrect her beloved grandmother. She has already let go of her grandmother's death, and the purpose of the action stems from the consensus of the whole family - her grandfather is ninety years old and may not be able to accept the death of his partner.For this reason, the whole family hid this from the old man for several months and kept deceiving him: Grandma was still lying in the hospital because of bad legs and feet.
Zhang Lin told Nanfeng Chuang that the whole family planned to continue this white lie, so she wanted to use AI to resurrect her grandmother's voice and talk to her grandfather.
's simple wish finally came true with 1,000 yuan - she handed over the 3-minute Cantonese conversation of her grandmother's life to the above-mentioned team, and the other party said one day later that the model was trained. Then, he asked Zhang Lin to imitate her grandmother's tone and voice habits to say a sentence. He then gave this sentence to the trained "grandma" and soon got the effect Zhang Lin wanted.
"When the audio came out for the first time, most of my relatives were shocked and startled," she recalled. "The timbre was very similar, about 95% (similarity)."
This is the majority of AI on the market. Resurrect the blogger’s main business. A Liang, an AI resurrected blogger, once told Nanfengchuang that there are three main categories of current business. The first is AI digital clone services. The second category is the speaking service, which essentially turns a photo into a video. “They actually just want their loved ones to say a blessing.” The third category is a text-only communication service for conversations with loved ones.
According to Nanfeng Window’s investigation, in the industry, the above-mentioned charges range from a few hundred yuan to ten thousand yuan.
But Nanfengchuang reporter found that this type of business focuses on information gaps. The first digital clone service, internationally, public software such as character ai and heygen ai have launched mature services.
takes heygen as an example. This project was created by a Chinese entrepreneurial team and will release instant avatar customized digital human services in 2023. Users can have a digital person with their own voice and appearance as long as they upload a front-facing video of themselves with clear speech and suitable lighting for more than two minutes.
This customized digital human service is charged on a membership basis and supports 25 languages (including Chinese). For a monthly fee of only US$49 (approximately 352 yuan), users can enjoy advanced customization services. In the future, by outputting relevant text, a digital person with your appearance can speak with your voice. Kevin, the ai voice blogger at
b station, introduced to Nanfengchuang that in the field of ai speech synthesis, there are many open source projects on the market. For example, gpt-sovits, a newly launched open source project in 2024, can easily and conveniently implement AI voice cloning.
Due to the existence of a large number of open source projects, Kevin believes that completing the resurrection of AI is almost zero cost for bloggers. "At most, it requires a few computer equipment equipped with GPUs." Coupled with the rise of AI cloud services, it is popular in the industry to rent GPU cloud services to achieve the computing power required to run AI models.
"You can spend a few dollars to rent an AI cloud service for an hour. If you are skilled, you can train five or six models in an hour. Basically, there is no cost to train a model." Kevin said.
is not as complicated as imagined.
kevin told Nanfengchuang that AI cloning speech technology has existed in the industry long before chatgpt, and is generally based on a technology called tts (text to speech). In China, the earliest commercial company doing TTS is today’s leading AI company, iFlytek.
The explosion of large models at the end of 2022 has aroused the interest of the technical community and commercial companies in AI voice cloning. Entrants are pushing this technology faster and faster.
"Now, you only need 1 minute of voice material, and after about 10 minutes of training, you can generate a clone model that is 90% similar to the original voice." Kevin introduced.
These advances benefit from algorithm innovation. He said that English software and models take less time and have more amazing results than Chinese ones. "Some open source English models have enabled users to upload 10 seconds of voice, and they can be cloned more closely."
03
The boundaries of technology
However, when people resurrected by AI flood the Internet in 2024, many people begin to feel uncomfortable.
's discomfort may be a criticism of the technology itself. On the other hand, "Compared with real people, AI's are still different," many people said.
A Weibo fan of Coco Lee told Nan Feng Chuang that when ai Coco Lee's videos flooded the Internet, she wanted to click in to watch the idol, but "opened it and listened to a few words and then closed it. It felt so fake."
This is also the limitation of many current AI cloning voice technologies. If you want to resurrect a familiar person and let him accompany you in your daily life, you will find that the technical threshold may still hinder many people.
kevin told Nanfengchuang that although AI can imitate timbre very well, it is still difficult to imitate the speaker's mouth habits and pauses. "For example, some people will take a breath when they speak. This is an effect that is difficult to achieve with AI open source projects." In addition, imitating human emotions, such as anger and disappointment, is a challenge to current AI cloning technology.
In March 2024, SenseTime, a leading domestic AI company, resurrected its founder Tang Xiaoou, who died of illness at the end of last year, at its annual meeting.
In the pre-edited video, AI Tang Xiaoou can drink water, joke, and said to the employees: "Everyone found it difficult last year, but I think that difficult things will always pass. In the end, we will be like " At the end of "The Long Season", a small train is sitting in the corn field, driving forward."
The SenseTime team revealed in an interview that in order to restore Tang Xiaoou, this 9-minute video used SenseTime's "Ruying" Technology, completed by the cooperation of several professional colleagues. It excerpted a large number of Tang Xiaoou's quotations during his lifetime. It took 2 months and five or six editions to come up with the current AI Tang Xiaoou.
In other words, ideal AI resurrection not only requires a large amount of high-quality voice materials from the living, but also a combination of details, time and technology.
What is even more difficult is to be like Bao Xiaobai, who enables real-time dialogue between the living and the digital people resurrected by AI. Three technologies are used: large language model, used to generate dialogue in real time; AI speech synthesis technology (tts), used to convert text into speech; AI digital human, used to display AI images on mobile phones.
When a Nanfengchuang reporter consulted a resurrected blogger named "AI Director Zhu" in the name of a consumer, he said that AI real-time dialogue requires a lot of computing power. "The hardware cost alone is four to five million, not including program development and maintenance." Therefore, the "factory director" said: Bloggers on the market are basically unable to communicate in real time.
He also added: "To do what you said, it takes six or seven masters to work for half a year."
The threshold of technology still exists, and the stars resurrected by AI are still far from real people. However, the video of AI resurrection has been spread more and more widely with the help of algorithms and traffic.
html On March 16, after Qiao Renliang’s parents and Gao Yixiang’s relatives expressed their discomfort with the resurrection of AI, many bloggers finally discovered that this move was suspected of infringing on portrait rights. According to Article 13 of my country's Civil Code, if the deceased's name, portrait, reputation, honor, privacy, remains, etc. are infringed upon, his spouse, children, parents, etc. have the right to request the perpetrator to bear civil liability in accordance with the law.In other words, relatives of celebrities who have been resurrected by AI have the right to pursue accountability from the video creator.
Many resurrected bloggers have since removed ai Qiao Renliang and ai Li Wei from the shelves...
But unstoppably, ai Qiao Renliang is still being forwarded and circulated many times on the Internet. People once again learned about this rapidly developing technology from the objections of celebrity parents.
"It's really scary. It has basically no cost. If criminals use this technology, they can train and imitate many people at the same time." Kevin, who studies AI, lamented.
Humanity once again stands at a fork in the road of choice. Technology is quickly ahead of everyone. Now, as Friedrich Dessauer, the founder of the philosophy of technology in the 20th century, described: "Modern technology changes the world and also witnesses its moral value beyond experience. Human beings have created technology, but its power is like a mountain or a river. , an ice age or a planet. It exceeds the various forces in the world."
At present, only the European Union has introduced relevant restraint plans.
In March 2024, the European Union passed the world's first "Artificial Intelligence Act." The highly cautious EU classifies deepfakes as "limited risk" AI systems.
This means that the management of deep synthesis technology focuses on prevention and does not touch upon punitive measures. The EU requires developers to implement technical safeguards to prevent the misuse of technology. For example, a watermark should be left on the AI output content.
-end-
editor | Xiangyou
editor on duty | Arshu
typesetting | Qiqi
Nanfeng Window New Media Produced by
Unauthorized reproduction is prohibited
Follow Nanfeng Window to see more exciting content